CVE-2026-18982 describes a privilege escalation vulnerability in the RHOAI training-operator component of Red Hat OpenShift AI. Users with edit or admin roles in Kubernetes namespaces can exploit this flaw by creating training jobs that allow impersonation of service accounts, unauthorized access to the host filesystem, and remote code execution. The root cause is the excessive permissions aggregated onto Kubernetes native ClusterRoles and the lack of restrictions on PodTemplateSpec passthrough. Red Hat has issued security advisories (RHSA-2026:53261 and RHSA-2026:53262) with updated RHOAI images (versions 2.25.10 and 3.4.3) to remediate the vulnerability.

An attacker with standard edit or admin privileges in any Kubernetes namespace can escalate their privileges to gain full control, including impersonating service accounts, accessing the host filesystem, and executing arbitrary code remotely. This can lead to complete compromise of the affected Kubernetes cluster and underlying host systems.

Red Hat has released updated RHOAI images (2.25.10 and 3.4.3) that address this vulnerability. Users should upgrade their Red Hat OpenShift AI clusters to these versions following the official Red Hat documentation to fully apply the errata update. No other specific mitigations are indicated by the vendor advisory.